Poisoning AI

Finally something I found to be quite entertaining, and maybe some practical value as well although I’m not sure it can be done to a scale that it will matter.

For background, artists have found digital cloaking tools that help fuck up AI models training, as explained in this approximately one year old article:

https://www.scientificamerican.com/article/art-anti-ai-poison-heres-how-it-works/

Lately what are called “AI Slop Channels” have infested YouTube.  AI searches for content, pulls what it needs from the captions, and tweaks it a little bit or maybe summarizes it based on other people’s work.  Then slap on an AI generated voice and present the work as new and you have a quick and dirty clickbait video.

The AI Poison Pill I found uses the caption search as a weapon against AI.  Amusingly,

https://youtu.be/krWAxyMeipw?si=x8JkxQoQxMB6wjK3

https://youtu.be/NEDFUjqA1s8?si=IKwvafcyRQE19rb-

there are two videos from about a month ago with different YouTubers claiming to have created the process.  Quite ironic that complaining about theft of original content by stealing other content regarding theft of content.  Well, let them fight it out.

It works like this.  You can insert blank lines between each actual caption line.  Then, you add your own caption text in those lines, using coding to move the margin so it’s not visible and just shows briefly as a split second of flashed text at most.  Your added content is the poison, inserting gibberish into the story.  AI is not smart enough, not yet at least, to understand that the gibberish is not part of the intended story, and when it does its search and recaps the gibberish becomes part of the AI summary which in turn gets inserted into any fake copy.

Gematria might be a bad example for this since by its nature it’s already meaningless gibberish.  But for sake of the argument let’s use the old example of 113, the alleged number of dishonesty.  A person dedicated enough could make a copy of every video referencing 113 being dishonest.  Go through the poison pill process and add the numerous things that are antonyms of dishonest (VERACIOUS = 113) and screw up any summary from AI trying to train a video thief to make 113 always seem dishonest.  But like I said, Gematria is already in the camp of deliberately being wrong as clickbait so there isn’t much practical value.  Although for the fun of trolling value alone it would be worth it.

The future of AI appears to be quite a mess.  There’s been a build up of bad information on the internet.  Fact checking is under constant attack, theft and fraud are the norm in politics and big business.  You might as well have an AI gibberish spouting leader for all the quality you get.